A systematic review of advances in spatiotemporal machine learning for road traffic accident prediction in developing countries

Low and middle-income countries (LMICs) account for 93% of global road traffic fatalities, yet spatiotemporal machine learning (ML) research remains structurally anchored in high-income contexts. This review synthesises a global evidence base while directing its analysis toward developing-country road safety: high-income-country studies (34.5%) are retained expressly to quantify, by contrast, the representation gap affecting low- and middle-income countries. This PRISMA 2020-compliant systematic review applies a dual mandatory inclusion criterion, requiring both a qualifying ML/deep learning algorithm and at least one explicit spatial and/or temporal analytical component, across 58 studies from 22 countries (2020–2026), with methodological quality assessed via PROBAST, marking its first application in road safety ML. Deep learning constituted the largest algorithm category (34.5%), with Random Forest the most frequently used individual algorithm (31.0%). The headline finding that 77.6% of studies employ combined spatiotemporal frameworks conceals substantial heterogeneity: only 17.2% of all studies (10/58) achieve true joint spatiotemporal learning, corresponding to 22.2% of studies within the combined-framework subset (10/45). PROBAST identified the Analysis domain as carrying the highest risk of bias 17.2%, primarily driven by unreported class imbalance handling and, critically, the complete absence of calibration reporting across all 58 studies, a universal limitation undermining model interpretability and deployment readiness. The Research Representation Ratio (RRR), applied here for the first time in road safety ML, suggests a pattern consistent with an inverse relationship between fatality burden and research intensity: high-income countries exhibit an RRR of 5.91 relative to their 7% fatality burden, while Sub-Saharan Africa shows an RRR of 0.27 relative to 26%. Bangladesh (RRR ≈ 1.15–1.72, roughly proportionate representation once WHO’s country-level fatality estimate is used) has no studies employing network-level GIS or temporal hotspot prediction. A five-priority national research agenda is proposed, addressing topology-aware spatial modeling, temporal hotspot prediction, environmental robustness testing, multi-source data integration, and resource-efficient hybrid modeling via transfer learning from proximate LMIC datasets.